Knowledge representation learning (KRL) aims to encode entities and relationships from the real world into low-dimensional, real-valued vectors. However, knowledge in specific domains is often distributed across multiple graphs, each complementing the others, posing challenges to current mainstream methods, which often fail to fully exploit the heterogeneity and complementarity of information across graphs. To address this issue, We propose MCKRL for multi-graph KRL, including two components to address the heterogeneity and complementarity of knowledge, respectively: (1) A hierarchical attention mechanism to convert diverse attributes into uniform-dimensional vectors, fusing them with attention from both attribute and source levels. (2) Introducing three channels (single graph, sub-graph, and combined graph), employing graph auto-encoders to learn entity embeddings in each channel, and combining them with graph-level attention for link prediction evaluation. Extensive experiments conducted on two representative knowledge graphs in biomedical research demonstrate the superior performance of MCKRL over the state-of-the-arts. The implementations areavailable at: https://github.com/wds-seu/MCKRL .